Stein Variational Belief Propagation for Multi-Robot Coordination

TL;DR

The paper tackles decentralized multi-robot coordination under uncertainty by modeling the robot swarm as a Markov Random Field and performing nonparametric marginal inference with Stein Variational Belief Propagation (SVBP). SVBP combines Stein Variational Gradient Descent with Particle Belief Propagation to maintain multi-modal marginal beliefs at each node, enabling parallelizable, GPU-friendly updates. Empirical results in both perception and planning tasks show SVBP better preserves true marginals and yields more robust planning, reducing deadlocks and producing smoother trajectories compared to PBP and GaBP, with successful real-robot demonstrations. The approach offers a scalable, decentralized alternative for high-dimensional, multi-agent coordination, robust to noise and asynchronous communication, while acknowledging computational scaling with graph degree and synchronization challenges as future work.

Abstract

Decentralized coordination for multi-robot systems involves planning in challenging, high-dimensional spaces. The planning problem is particularly challenging in the presence of obstacles and different sources of uncertainty such as inaccurate dynamic models and sensor noise. In this paper, we introduce Stein Variational Belief Propagation (SVBP), a novel algorithm for performing inference over nonparametric marginal distributions of nodes in a graph. We apply SVBP to multi-robot coordination by modelling a robot swarm as a graphical model and performing inference for each robot. We demonstrate our algorithm on a simulated multi-robot perception task, and on a multi-robot planning task within a Model-Predictive Control (MPC) framework, on both simulated and real-world mobile robots. Our experiments show that SVBP represents multi-modal distributions better than sampling-based or Gaussian baselines, resulting in improved performance on perception and planning tasks. Furthermore, we show that SVBP's ability to represent diverse trajectories for decentralized multi-robot planning makes it less prone to deadlock scenarios than leading baselines.

Figures (10)

• Figure 1: Stein Variational Belief Propagation (SVBP) computes marginal trajectory distributions for each robot in a multi-robot system. SVBP represents the relationships between robots as a Markov Random Field (a) and maintains multi-modal distributions over each robot trajectory (b). Example final trajectories for each robot are shown in (c).
• Figure 2: SVBP better represents the underlying distribution, avoiding mode collapse. (a) Graphical model of the multi-robot perception problem. The position of each node is denoted $x_i$, and the corresponding observation is denoted $z_i$. (b) The approximate true marginals for the graph in (a) and the observation shown in (c, d). Qualitative results for SVBP (c) and PBP (d) at the final iteration ($k$). The red lines represent the true position of the nodes, and the colored 'x' markers represent the maximum likelihood estimate for each node. Lower-weighted particles are shown with lower transparency. The distributions represent the noisy observations for each node of the corresponding color. Best viewed in color.
• Figure 3: Average error for each node estimate for multi-robot localization. Results are shown for varying levels of noise, corresponding to the number of noisy components added to the observation.
• Figure 4: The average Maximum Mean Discrepency (MMD) between the samples from the true marginal distribution and the particle sets from SVBP and PBP. Both methods use 50 particles.
• Figure 5: Average error for each node estimate for different numbers of particles. The solid lines correspond to experiments runs with noise added to the observation. The dashed lines correspond to experiments with no noise added to the observation.